Writing

Essays, research and market maps

Where I think AI goes next and where the value lands: essays with a stated thesis, ecosystem maps, seed-stage deal memos, and forecasts that say in advance what would prove them wrong. Jump to any section below.

Essays

What comes next in AI, and where the value goes

Long-form arguments with a stated thesis, the evidence for it, and the signals that would prove it wrong.

Oct 20269 min readAIVenture

From data to crypto to AI: three waves and what each one taught investors

Cloud, crypto and generative AI each bent a cost curve. Here is where the money went each time, and what that says about AI.

ThesisOver the next five years, AI value shifts from compute and model suppliers toward application companies that own a specific, high stakes customer workload, because the cost of model capability keeps falling. Read the essay
Oct 202610 min readAIVenture

Vertical or horizontal: where AI applications will win

Horizontal AI wins the headlines. I think vertical AI built around a liability bearing workflow builds more durable companies, and here is how I underwrite both.

ThesisThe most durable AI application companies of the next five years will be vertical ones that own a liability bearing workflow and price against labor, while most horizontal application value goes to model labs and suite incumbents, with coding as the main exception. Read the essay
Oct 20269 min readAIRobotics

From text to bodies: the move to embodied AI

Robot foundation models are improving fast, but outside robotaxis the deployments are still measured in hours per robot, and that gap is where I would invest.

ThesisOver the next five years, the durable returns in embodied AI go to companies that own deployed robot hours in narrow, high-labor-cost jobs and to the component, simulation and safety layers beneath them, not to general-purpose humanoid makers priced on demos. Read the essay
Oct 20269 min readBiocomputingDeep tech

Hardware, firmware, software, fluidware

Fluidware is my word for computing on living neurons: the science is real but small, and the commercial case is still a long bet.

ThesisFor the next decade, fluidware makes money as a research and drug-discovery tool, and the investable layer is the infrastructure that keeps neurons alive, connected and measured, not biological computing as a rival to silicon. Read the essay
Market maps

Who builds what, layer by layer

Ecosystem maps for the AI stack, embodied AI and biocomputing. Each company placement was checked against what the company actually ships.

The AI stack, October 2026Read it bottom to top: each layer buys from the one below it and sells to the one above it.
Vertical applicationsDomain workflows in regulated or expert industries.
LegalResearch, drafting and case work for lawyersHarveyLegoraEvenUp
HealthcareClinical notes, coding, evidence search, patient voice agentsAbridgeAmbience HealthcareOpenEvidenceHippocratic AI
Financial servicesResearch and deal work for bankers and investorsRogoHebbiaAlphaSense
Drug discoveryModels that design or find new medicinesIsomorphic LabsRecursionChai Discovery
Horizontal applicationsTools any team can buy, regardless of industry.
Coding agentsWrite, test and ship code from instructionsCursor (Anysphere)CognitionReplitGitHub Copilot
Enterprise search and assistantsAnswers and agents over company dataGleanMicrosoft 365 CopilotChatGPT Enterprise (OpenAI)
Customer service agentsResolve support conversations across chat, email, voiceSierraDecagonIntercom (Fin)
Tooling and orchestrationWhat teams use to turn a model into a product.
Data and retrievalGoverned data, training data and vector retrievalDatabricksSnowflakeScale AIPinecone
Evals and observabilityTrace, score and debug agents in productionBraintrustArize AILangChain (LangSmith)
Agent frameworksBuild multi-step and multi-agent workflowsLangChain (LangGraph)LlamaIndexCrewAI
Foundation modelsGeneral models others build on.
Frontier labsLargest closed models, sold by API and appsOpenAIAnthropicGoogle DeepMindxAI
Open-weight modelsWeights you can download, tune and hostMeta (Llama)Mistral AIDeepSeekAlibaba (Qwen)Moonshot AI (Kimi)
Specialist and enterprise modelsVoice, image, video and private enterprise deploymentsCohereElevenLabsBlack Forest LabsRunway
Cloud and infrastructureWhere models are trained and served.
HyperscalersGeneral clouds with large AI capacityAWSMicrosoft AzureGoogle CloudOracle Cloud Infrastructure
GPU cloudsSpecialist clouds built around AI acceleratorsCoreWeaveNebiusCrusoe
Inference platformsServe and fine-tune open models at scaleTogether AIFireworks AIBaseten
Compute and chipsThe layer that captures most of the spend today.
AcceleratorsGPUs and custom silicon for training and inferenceNVIDIAAMDGoogle (TPU)AWS (Trainium)Cerebras
Foundry and memoryLeading-edge fabs and high-bandwidth memoryTSMCSK hynixSamsungMicron
NetworkingSwitching and custom silicon linking GPU clustersBroadcomArista NetworksMarvell

As of Oct 2026

Sources: NVIDIA Q2 fiscal 2027 results Cerebras CoreWeave Nebius Crusoe Together AI Fireworks AI Baseten Mistral AI Cohere ElevenLabs Black Forest Labs Databricks Snowflake Q2 fiscal 2027 results Scale AI Pinecone Braintrust Arize AI LangChain LlamaIndex CrewAI Cursor Cognition Replit Glean Sierra Decagon Fin by Intercom Harvey funding, Sept 2026 Legora EvenUp Abridge press releases Ambience Healthcare OpenEvidence press releases Hippocratic AI Rogo Hebbia AlphaSense Isomorphic Labs Recursion Chai Discovery

Deal memos

Companies before the breakout

Seed-stage deal memos, one for each of my big bets: the round, the team, what I would have to believe and the questions I would ask first. They live on the Venture page.

01Seed · $6.5MWould invest with conditions

Avatar Robotics

Remotely operated humanoids that do warehouse work today, sold by the hour, with every shift recorded as training data for the autonomous version.

Led by AlleyCorpAug 2026 Big bet: General-purpose robots Read the deal memo
02Seed · $4.3MWould invest

Ember AI

An AI revenue team for specialty physician groups: it checks every claim before it goes to the insurer, then fights the denials that still come back.

Led by Nexus Venture PartnersNov 2025 Big bet: Services as software Read the deal memo
03Seed · €5MWould invest with conditions

Hyperscale Power

A solid-state transformer that takes grid power straight to the 800-volt DC that AI racks now want, replacing equipment with 18-to-24-month lead times.

Led by World Fund and Vsquared VenturesMar 2026 Big bet: Power for AI Read the deal memo
04Seed · $15MTrack closely

Axiom Bio

Its own human liver-cell lab plus a model that predicts which drugs will injure the liver, pitched as a replacement for animal toxicity studies.

Led by Amplify PartnersApr 2025 Big bet: AI-designed medicines Read the deal memo
05Seed · $10.3MWould invest

Critical Materials Group

Modular plants that make military explosives, starting with C-4, built to add capacity next to the few aging government plants the US depends on.

Led by Overmatch VenturesSep 2026 Big bet: Defense and reindustrialization Read the deal memo
06Seed · $4.5MTrack closely

UnblockPay

An API that turns stablecoins into local money and back across Latin American bank rails, sold to brokers, small banks and fintechs.

Led by PreludeMar 2026 Big bet: Stablecoins and agent payments Read the deal memo

Nothing here is investment advice: these are my own notes on public information, written to sharpen judgment, not to recommend buying or selling anything.

Research

Forecasts and working papers

Each piece states in advance what would prove it wrong: calibrated probabilities, pre-registered hypotheses, or an explicit list of what the method cannot see.

Jun 2026

The Chokepoint Decade

Forecasting the collision of AI and modern mercantilism, 2026 to 2036

AI and mercantilism as one competition for five scarce inputs: capital, energy, critical minerals, semiconductors and skilled labor. 24 binary forecasts with objective resolution criteria, calibrated from 0.40 to 0.88.

Bridgewater Associates × Global Citizen, Forecasting the Future Challenge

2026

Measuring Buyer Intent from Costly Public Signals

Working paper

Weights every public buyer action by what it cost the actor to produce. Intent in log-odds with Good's weight of evidence, a hazard model so silence counts, and an explicit contract for what the instrument cannot see.

Working paper

Jul 2026

Predicting Fan Engagement Under Sparse Signals

A pre-registered cold-start design for the 2026 FIFA World Cup

Extends latent-attrition models into the cold-start case with a graph-structured prior. Seven falsifiable hypotheses registered before any data was fit.

Working paper and reproducible pipeline

Jul 2025

Forecasting the AI HealthTech Decade

Power, policy and wellbeing, 2025 to 2035

Health data and clinical AI as sovereign infrastructure. 20 calibrated forecasts and a macro feedback model tested against five country cases.

Independent research paper

Abstract

The ROI of Superfans

Lifetime fan value from social graphs and engagement heatmaps

1.2 million fan profiles and 150 million touchpoints, graph embeddings and survival regression, 88.4% out-of-sample accuracy against an RFM baseline.

MIT Sloan Sports Analytics Conference, past first-round review

Published on the products I build

Research reports, posts, guides and a manifesto I wrote for Searchalong and Keystone. Each opens on the product's own site.

Send me what you are working on.

A product idea, a team that needs another builder, a competition that needs a judge. Tell me what it is and I will say straight away whether I am the right person.

Book a 30-minute call LinkedIn
↑ ↓ to moveEnter to openEsc to close